Skip to content

Source paper

Helios: Real Real-Time Long Video Generation Model

Shenghai Yuan, Yuanyang Yin, Zongjian Li and 3 more

Shenghai Yuan, Yuanyang Yin, Zongjian Li, Xinwei Huang, Xiao Yang, Li Yuan

arXiv 2603.04379 · 2026117 upvotes on Hugging Face

Research ideas from this paper (3)

Generated by AI. The paper’s authors did not write these.

AI summary

Helios introduces a 14-billion parameter autoregressive video generation model capable of synthesizing minute-scale videos at 19.5 FPS on a single H100 GPU. The authors achieve this by adapting a pre-trained bidirectional transformer with novel token compression techniques, drift-mitigation strategies that avoid expensive training-time rollouts, and a hierarchical adversarial distillation pipeline.

How it works (AI summary)

The model transforms the Wan-2.1-14B DiT into an autoregressive generator using 'Unified History Injection' and 'Easy Anti-Drifting' techniques (Relative RoPE, First-Frame Anchor). It achieves real-time speed via 'Deep Compression Flow,' which utilizes Multi-Term Memory Patchification to compress history tokens and a Pyramid Unified Predictor Corrector to reduce computation in early noisy sampling stages, followed by Adversarial Hierarchical Distillation to reduce inference to 3 steps.